Ubuntu Image 20150130 OpenCL

[nickoppen]

I'm wondering what the status is of the current Ubuntu released Jan 30 2015. There was a thread requesting Beta testing but that seems to have disappeared.

I've been testing the OpenCL.

As released with version 1.6.0 of the Brown Deer OpenCL compiler, the in-build tests mostly worked. By mostly, I mean that they seemed to crash when the ARM was busy (i.e. boinc running in the back ground) but when they were the only thing going they mostly worked.

With v. 1.6.0 none of my programs that I'd gotten working on the previous image worked. The all ended up either stalled or locking up the whole machine.

The upgrade to 1.6.1 was a complete failure with libelf missing.

The upgrade to 1.6.2 went better. I can now compile and get an executable compiled from source. I've gotten most of my executables to run with some tweaking of linker parameters. There still seems to be a problem with libclelf. When I try and link the kernel executables to an elf file I get a "libclelf.so not found" error.

The big surprise with 1.6.2 is that now the test process seems to lock up every time regardless of what is happening. Mostly it gets through the first test but locks up on the second. Sometimes it gets 7 or so done.

I've also reinstalled and recompiled the riecoin miner (fastrie/xptMiner). That mostly has a sieve error after 1:12.

On the plus side the reboot success rate is much better. Not 100%, probably more like 80%.

I hope this helps.

nick

[nickoppen]

I'm going back to the previous image. While the performance improvements are welcome, the lack of reliability of the kernel calls means that it is not a viable platform for OpenCL development.

How I got to this conclusion was:

Using the 1.6.2 version of the compiler I got my data passing test compiled. At first it didn't run at all. It locked the whole machine up on the first forka command. This was from the same source code that worked consistently on the old image.

Assuming that I needed to tweak my code, my first investigation was to wrap all of the kernel source with an if(0) { ... } so that none of my code would be executed. The intention was to "unblock" the rest of the kernel bit-by-bit to see where it locked up.

The result of this was that sometimes the kernels returned and sometimes they didn't. It varied between all 16 calls to all four kernels returning to it locking up on the first call.

This demonstrates to me that my code was not causing the problem. The behaviour was very much like the coprthr tests. Sometimes they work and sometimes not.

Sorry I can't give any more details. The gap between the forka and the kernel is a black box to me.

[olajep]

[nickoppen]

OK

I'll have another try.

Come to think of it, my fan started playing up when I switched over to the new image. I thought that the thermal deamon was doing something clever with the fan but having read the description on github I see that it is not that clever.

[olajep]

[nickoppen]

I think you are right.

When I reverted back to the old image and noticed that the fan did not start I had a look at the temperature with xtemp. It was sitting around 68 - 69 degree range with boinc running so I can imagine that it didn't need to do much more for the daemon to kick in. That's probably why the riecoin miner fell over straight away.

Running the coprthr tests without boinc running it took a minute or two to lock up. Probably about the amount of time it took for the extra activity to heat things up. I ran the whole suite about 10 times consecutively last night with the fan on and there were no lockups.

I've got the large new version heat sink attached and there is good air flow over it.

It's not that hot in Sydney. We're actually getting cooler here now, heading towards our horrendous winter temperatures of around 15 degrees C (that's 15 above zero) and my wife and daughter are complaining about the cold already. This means that the heating is getting used quite a bit (insane, I know). My room where the parallella lives is hotter than the rest of the house because of my windows machine. When I came in this evening I had a look at the temperature using the shell script that you posted. It was a bit above 70 degrees. Now that I've got the door open it sitting around 66. I'll try and keep the door open but is it possible to set the temperature threshold to a slightly higher figure (say 75C)?